Research map / Point clouds / Registration and matching
Coupled Laplacian Eigenmaps for Locally-Aware 3D Rigid Point Cloud Matching
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Coupled Laplacian, a training-free graph Laplacian eigenmap operator that aligns spectral embeddings of registered point clouds via cross-edges, and applies it to MVTec 3D-AD anomaly localization and a new bone side estimation benchmark.
Problem
Rigid 3D point cloud matching that is sensitive to fine local differences; existing spectral descriptors suffer from eigenvalue order and sign ambiguity, which breaks matching on symmetric shapes like bones.
Approach
Merge target and registered source point-cloud k-NN graphs with artificial cross-edges between sampled correspondences, so the eigendecomposition of the single coupled Laplacian yields automatically aligned eigenmaps for each shape; compare them via Grassmann distance (global) or cosine distance (point-wise).
Key results
- Ours100 achieves mean PRO AUC 0.736 and Ours200 0.748 on MVTec 3D-AD, versus 0.703 for the best Euclidean baseline and 0.627 for Hist200.
- Ours10 reaches 82.39% mean accuracy on human bone side estimation, versus 80.72% for Hist20 and 77.35% for Chamfer.
- Registration failure rate for femur was 6.84 ± 7.82% with 96.67 ± 11.30% BSE accuracy on failures, and hip had 0% failures.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Coupled Laplacian eigenmaps instead of Euclidean point distances (ablation) | point-wise anomaly localization · MVTec 3D-AD | area under PRO curve (limit 0.3), mean: 0.703 -> 0.748 | Helped |
| use Coupled Laplacian eigenmaps instead of histogram-based eigenfunction matching | BSE · Human BSE | accuracy, mean: 80.72 -> 82.39 | Helped |
| use 200 eigenmaps instead of 100 (ablation) | Coupled Laplacian anomaly localization · MVTec 3D-AD | area under PRO curve (limit 0.3), mean: 0.736 -> 0.748 | Helped |
| use 20 eigenmaps instead of 10 (ablation) | Coupled Laplacian BSE · Human BSE | accuracy, mean: 82.39 -> 81.83 | Hurt |
| compare aligned Coupled Laplacian eigenmaps rather than Euclidean coordinates (ablation) | point-wise anomaly scoring · MVTec 3D-AD | anomaly localization | Helped |
| increase post-registration rotation and translation noise (ablation) | Coupled Laplacian BSE · Human BSE | accuracy drop: 2.5 / 5 / 8.16 - 5 / 5 / 15.13 - 7.5 / 10 / 18.16 | Hurt |
Surprising findings
- Point-wise similarity on unaligned eigenmaps (GPS, Hist) performs worse than plain Euclidean distance between registered points, making spectral embeddings without alignment worthless for this task.
- Eigenmap-based matching on bones with high symmetry fails often; the method's fallback is sensitive to the fraction of cross-edges used.
Limitations
- Usage is limited to tasks where correspondences can be defined via rigid or affine registration; the landmark-free approach relies on registration quality.
- Registration failures and rotation/translation noise degrade accuracy, and high-symmetry bones (femur) are more prone to failure.
- Performance depends on the number of eigenmaps and the fraction of cross-edges, requiring tuning.
Experimental setup
Datasets: MVTec 3D-AD Fisher et al. bone surface models SSM-Tibia (Keast et al.) ICL bone models (Nolte et al.) internal sheep bone dataset
Benchmarks: MVTec 3D-AD BSE benchmark
Metrics: area under PRO curve (integration limit 0.3) classification accuracy
Baselines: Euclidean distance GPS Hist (eigenfunction histogram matching) MAP FPFH Chamfer distance Hausdorff distance GAN AE VM
Builds on: Laplacian eigenmaps Global Point Signature Coherent Point Drift RANSAC Laplacian Canonization (MAP)
Closest papers by meaning
- Volumetric Functional Maps (CVPR 2026 · similarity 0.85)
- Spectral Meets Spatial: Harmonising 3D Shape Matching and Interpolation (CVPR 2024 · similarity 0.85)
- APML: Adaptive Probabilistic Matching Loss for Robust 3D Point Cloud Reconstruction (NeurIPS 2025 · similarity 0.85)
- SPEAL: Skeletal Prior Embedded Attention Learning for Cross-Source Point Cloud Registration (AAAI 2024 · similarity 0.84)
- X-3D: Explicit 3D Structure Modeling for Point Cloud Recognition (CVPR 2024 · similarity 0.84)
- Hybrid Functional Maps for Crease-Aware Non-Isometric Shape Matching (CVPR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.